Optimizing Air Pollution Forecasting Models Through Knowledge Distillation: A Novel GCN and TRANS_GRU Methodology for Indian Cities

As air pollution remains a pressing issue affecting environmental and public health, precise forecasting and monitoring are essential. Machine learning and deep learning models are increasingly applied in air pollution studies, using their advanced capabilities to decode intricate pollutant data. A review of the existing literature on air quality forecasting reveals significant gaps: first, a lack of focus on identifying dominant pollutants for specific dates and cities, and second, the widespread use of models with excessively high computational demands. To address these challenges, we introduced the Graph Convolutional Networks (GCNs) for finding single and multi-dominant pollutants, and Transformer_Gated Recurrent Unit (TRANS_GRU) hybrid deep learning model for improved accuracy in the prediction of short and long-term pollutants levels subsequently applied Knowledge distillation (KD) for efficient, lightweight modelling. This approach demonstrates superior performance over traditional models such as long-short-term memory (LSTM) and LSTM_BiGRU (LSTM_Bidirectional Gated Recurrent Unit) in various metrics. The TRANS_GRU model achieves improvements of 2.36% in the R2 score, 15.23% in the mean squared error (MSE), 19.12% in the mean absolute error (MAE), and 7.93% in the root mean squared error (RMSE) compared to LSTM. Against LSTM_BiGRU, shows enhancements of 4.71% in the R2 score, 27.64% in the MSE, 33.19% in the MAE, and 14.86% in the RMSE. These advances not only address existing research gaps, but also equip environmental scientists and public health professionals with more effective tools to manage air quality every day.